Iris: How to Climb to the Search Frontier

Patrick Rho · AI Research

I read Iris: Climbing to the Search Frontier, published on arXiv as arXiv:2609.04304. The central claim is that a search agent can reach the frontier of search benchmarks with a single agent alone when the data pipeline and the training recipe are designed together and SFT and reinforcement learning are alternated on the way up. The original paper is linked under the title.

This post follows that climb at walking pace. The first half looks at how questions are built in reverse, how clues are masked, and how only trainable questions are kept. The second half looks at how trajectories are refined, how reinforcement learning runs against live search, how overlong rollouts are salvaged, and how the two stages alternate upward. It closes by reading the numbers, examining the single-agent choice, and listing the questions that the released recipe leaves behind. Numbers are summit photos and process is the route map. This post spends most of its time on the route map.

Why the search frontier matters now

Search is one of the oldest testing grounds for language models. The line between what a model knows and what it does not know shows up clearly in search. A model that sounds confident with no books open starts to wobble once sources are demanded. Attach search and answer quality changes, but new decisions appear. The system must decide what to search, how to read the results, and when to stop and answer. The phrase search agent bundles these three decisions into one name. It means a system that runs the loop of querying, reading, and choosing its next move on its own.

Running that loop well turns out to be harder than it looks. Questions answered by a single search are largely solved. What remains are questions that must be pieced together across pages. When the bridge from one page to the next breaks, the agent wanders. A slightly wrong query drops it into a swamp of irrelevant documents. Reading too much overflows the context and reading too little misses the clue. These failures do not vanish just by scaling the model. The data that teaches behavior and the reward that shapes behavior must change together.

That is where Iris gets interesting. The striking fact is not a declaration of a larger model but the release of a whole climbing route. The data pipeline and the training recipe arrive together. Usually the score table comes first and the method leaks out later in fragments. Here the order is reversed. How to build questions, how to select them, how to refine trajectories, and how to assign reward arrive as one bundle. Scale is reported alongside. Iris-mini sits at the 35B-A3B scale and Iris-pro at the 397B-A17B scale. The fact that the small one and the large one climbed the same route matters. It suggests the recipe is a procedure that spans scales rather than a trick tied to one size.

When my team reviews a new search study, we also look at the training process before the score table. We ask what questions it was fed, which behaviors were rewarded, and where the training pushed and where it held back. Scores are summit photos. Process is the route map. Summit photos impress but do not help the next climb. A map lets the next team follow. Iris reads like a study that invested in mapmaking. So this post also saves the score discussion for later and gives the early sections to data and training.

Building questions in reverse from hyperlink structure

The Iris data pipeline starts from an unusual place. It does not write questions first and find answers later. It starts from the connection structure of web pages and builds questions in reverse. It picks one seed page, gathers the outlinks that the page points to, distills an entity graph from that bundle, and authors multi-hop chains that tie several entities together. The chain becomes the skeleton of the question. The work writes the question with the answer and the evidence already in hand. It resembles an examiner who fixes the answer and the sources first and then writes the test item.

This order is attractive because the evidential links are guaranteed from the start. When questions come first, evidence hunting becomes a later chore. A question can sound plausible while its support sits on no single page or scatters across pages that never connect. Reverse construction flips the sequence. A bundle of connected pages comes first and the question is authored on top of it. The risk of a broken path to the answer shrinks. Hyperlinks already form trails laid by human hands. A bundle of documents that point to each other is a ready warehouse of reasoning chains.

Look one step closer and the simple phrase of seed page plus outlinks hides many design choices. Which pages qualify as seeds, how many steps to follow the links, and what to keep or drop in the entity graph must all be fixed. A seed that is too famous makes questions too easy, because the agent can jump to a memorized answer. Links spread too wide make the graph messy, with unrelated entities slipping in and chains turning forced. Links kept too narrow make questions monotonous, with only one hop or two hop items rolling out. Good reverse construction lives somewhere in between. It needs the touch that pulls out chains that are varied yet genuinely connected.

There is also a difference between listing entities and authoring a chain. A list needs no order. A chain lives or dies by order. The first link must point at the second, and the second must open the third. Each step must give the agent a reason to search again. If one search resolves everything, it is not a chain. A genuine multi-hop question has a structure where each link calls for the next search. The emphasis on multi-hop chains reads as a declaration of intent to manufacture such structures on purpose rather than waiting for hard questions to appear by accident.

Blocking string matching by rewriting every non-answer entity

The device that follows reverse construction shows the craft of this work. Every non-answer entity is rewritten into a descriptive reference. Names are not given as clues in their original form but replaced with descriptions that point at them. A person, place, or title appears as a descriptive phrase rather than a literal string. Then no clue can be resolved by copying a phrase into a search box. It is a barrier that shuts the shortcut of string matching.

Why is such a barrier needed. Evaluation of search agents is vulnerable to cheating. When a question carries a distinctive proper noun that leads straight to the answer, the agent gets good at copying rather than reasoning. It slices the question, pastes the slices into the search box, and the answer pops out. Scores rise while skill stays flat. Benchmarks mean to measure composition rather than copying, the ability to read several documents and connect their meanings toward an answer. Descriptive rewriting blocks the copying shortcut and leaves only the composition path. It does not say take a detour. It says travel properly.

Applying this rule in practice is far from trivial. The team must decide what counts as an entity, how to separate answer entities from the rest, and how detailed each description should be. A description that is too detailed merely paraphrases the name, and the agent can search the hint text and fall back to copying. A description that is too vague breaks the question, leaving a sentence that no human reader could resolve. A good descriptive reference sits somewhere in the middle. It reads clearly in meaning yet yields no direct hit when pasted into a search box. It keeps meaning open while keeping strings closed.

This passage reminded me of an old worry in evaluation design. There are two ways to make a question harder. One demands more knowledge and the other demands more action. The first favors whoever memorized more. The second favors whoever searches and connects better. Search agents should compete on the second. Descriptive rewriting shuts the shortcut of the first kind and leaves the road of the second kind. One rule changes the character of the whole dataset. From the same bundle of pages, rewritten questions teach different behavior. The agent learns not what to memorize but how to find.

Keeping only trainable questions between closed book and open evidence

Once questions are built, they must be selected. Iris states a sharp rule. Only questions that a reference model fails closed book yet solves once supporting evidence is supplied are admitted to training. Both conditions must hold. Questions solvable from memory are dropped. Questions unsolvable even with evidence are dropped. Only the band in between is kept, where the model cannot answer alone but can answer with materials. The judgment is that learning lives exactly in that band.

The rule is sharp because it blocks two failures at once. The first failure is feeding questions that are too easy. When the model already knows the answer, repetition adds no behavior. The agent learns to pretend to search with no reason to search. Time and compute drain away. The second failure is feeding questions that are too hard. A question that resists even with evidence is a wall for the current model. Repeated headbutting lengthens trajectories without reward. Trainable questions sit in the narrow band between. They are questions where search and reading make the difference.

The phrase reference model deserves a pause. Who serves as the yardstick changes which questions survive. A yardstick that is too weak labels easy questions as hard. A yardstick that is too strong discards trainable questions as trivial. A good reference model should stand at a proper distance from the trainee, slightly ahead or side by side. Too far ahead makes the bar cruel and too far behind makes it lax. The presence of a reference model reads as a choice to handle selection as procedure rather than feel. Instead of humans reading and picking each item, an automatic sieve of model success patterns does the filtering.

There is another layer in the phrase once evidence is supplied. Solvability with supporting evidence means the answer truly lies inside the sources. If the answer lies outside, evidence cannot help. So the condition also audits the coherence of each item. It tests whether the reverse-built chain truly connects and whether descriptive rewriting broke anything. Failure with evidence hints that some link snapped or some description blurred past recognition. Selection by difficulty is therefore also quality control for the whole pipeline. The loop that builds in front and verifies behind turns one full circle here.

Refining trajectories and turns before supervised fine tuning

Selected questions produce trajectories when the agent runs on them. Each record strings together searching, reading, searching again, and answering. Iris filters these trajectories at two levels before supervised fine tuning. Trajectory level filtering asks whether the whole record is worth teaching. Turn level filtering asks whether each single move inside deserves teaching. A record with the right direction keeps its good steps while its bad steps are repaired. A record with the wrong direction is discarded whole.

Two levels are needed because agent records are long and lumpy. Some trajectories reach the right answer through a mess. A lucky query lands or a last minute guess saves the run. Teaching such a record whole teaches the bad habits along with the answer. Other trajectories miss the answer despite excellent searching in the middle. Good exploration slips at the final judgment. Discarding such a record whole throws away good behavior with the failure. The two level filter separates the two. The trajectory level judges direction and the turn level judges footsteps. The intent is to teach only good steps inside well directed records.

Anyone who has opened raw agent logs knows the feeling. Records are messy and paths to the same answer vary wildly. One trajectory rewrites its query three times and arrives. Another reaches the neighborhood on the first search and then wastes ten redundant checks. Both read as correct on the surface. Their teaching value differs. The first demonstrates repair and the second demonstrates drift. Turn level filtering is the magnifier that separates them. It tells which search created progress and which reading spent time. A sharper magnifier raises the density of supervised learning. The same question count teaches more craft.

This stage also levels the starting line for the reinforcement learning that follows. Supervised tuning teaches the basic gait. A crooked gait makes reinforcement learning wander, chasing reward into strange dances. A solid gait lets reinforcement learning focus on direction. The two level filter paves the starting ground by removing wobbly stones and keeping only stones worth stepping on. To use the climbing metaphor, it maintains base camp. Base camp work looks less glamorous than the summit push, yet climbs are often decided there.

Reinforcement learning against live search with judge and summarizer inside

With the gait steadied, the climb enters real terrain. The policy is optimized by reinforcement learning against live search. It learns by throwing queries at living search rather than circling inside a precollected bundle. The reward judge and the observation summarizer are served inside the training cluster. Judgment and summarization run nearby instead of being requested from outside services each time.

The design matters for speed and stability. Live search is slow and jittery. Long outside delays tie learning steps to waiting. A reward decision placed outside wobbles or stalls the whole run whenever calls clog. Housing the judge and the summarizer inside the cluster keeps the loop fast and steady. The agent searches and reads, the summarizer tidies the observation, and the judge assigns reward. The circle spins quickly inside the cluster. Faster spins mean more trial and error per hour. Since trial and error is the fuel of learning, spin speed is learning speed.

The phrase reward judge deserves a pause. What to reward in a search agent is the heart of the design. Rewarding only final correctness can corrupt the process. Rewarding speed alone teaches skipping the reading. Rewarding long reading rewards wandering. A good judge weighs answer and process together, measuring correctness, evidential continuity, and wasteful steps in some imagined blend. Keeping the judge inside the training cluster reads as a choice to hold the yardstick close rather than leaving it to outside moods. It declares that the team will grip the measure and keep it steady.

The observation summarizer deserves equal thought. Search output is long and messy. Feeding it raw to the policy overflows context and slows learning. The summarizer trims observations before handing them over. What to keep and what to drop is the skill. Dropping too much erases clues and dropping too little leaves noise. Placing the summarizer inside the cluster pulls the tidying work into the learning loop rather than leaving it to outside chance. It manages reading quality as procedure. Of the two legs of search and reading, it strengthens the reading leg.

Salvaging long rollouts by resuming from the committed prefix

Reinforcement learning against live search often runs long. Searches chain into more searches and readings lengthen until trajectories sprawl. Iris handles overlong rollouts by interrupting them at the request level and resuming from their committed prefix. It does not discard the whole run but keeps the salvageable front and continues from there. In climbing terms, it halts an overambitious solo push at an intermediate camp, keeps the fixed ropes, and continues the attempt from secured ground.

The device is practical because it wastes less compute. Discarding a long rollout whole also discards good early exploration. Often the first steps were sound and only the tail drifted. Fixing the front and continuing from it saves the good search while cutting the bad tail. Compute waste falls. The density of the learning signal rises. It is salvage work that lifts the rescuable front out of a pile of failed records.

Look closer and the phrase committed prefix raises a design question. Where to draw the line of commitment is the craft. A line drawn too short restarts from scratch every time. A line drawn too long cements bad steps as trusted ground. A good cutting point depends on a yardstick for the front steps, whether searches created progress and whether readings left clues. A dull knife discards what should be saved and saves what should be cut. A sharp knife raises efficiency. The detail looks minor, yet in large scale reinforcement learning such minor efficiencies decide outcomes. The larger the cluster and the longer the runs, the more the interrupt and resume design sets the overall pace.

The phrase at the request level also invites thought. It suggests managing one agent run not as a single block but as splittable requests. Seen whole, a long run is one failure. Split apart, it divides into a front worth keeping and a tail worth cutting. Changing the unit of view turns waste into resource. It matches the practical sense of teams that husband compute. When compute is dear, the art of salvaging matters more than the art of discarding. This device moves that sense into procedure.

Climbing by alternating SFT and RL on the hardest and most efficient rollouts

Now the route reaches its crux, the procedure called SFT-RL climbing. Supervised tuning and reinforcement learning alternate upward instead of finishing in one stage. The materials for each round are specific, the hardest solved rollouts and the most efficient rollouts. The hard ones carry the most lessons and the efficient ones model the path to copy.

Why alternate. Pushing one stage alone always stalls somewhere. Supervised tuning alone settles into mimicking exemplars and stumbles in unseen situations. Reinforcement learning alone gets trapped in short term reward tricks that lift scores thinly. Alternation lets each cover the other. Supervised tuning steadies the basics and reinforcement learning explores above them. When reinforcement learning finds a new path, supervised tuning writes it into the basics. It resembles a mountain walk that alternates ascents with flat stretches, gaining height on the climbs and recovering breath on the flats.

The choice of materials catches the eye. Using the hardest solved records means refusing to repeat easy wins. A hard solve means the agent stumbled before reaching the answer. Such records contain both the stumble and the escape, where it stuck and how it broke out. Teaching them passes on escape craft. The most efficient records show paths with no wasted steps, where to skip and where to focus. Teaching them passes on shortcuts. Together they teach escape and shortcut at once. Hardship stories and model answers serve as textbooks side by side.

Seen over time, the procedure climbs under its own power. Hard successes become textbooks for the next round, and the policy hardened by those textbooks finds harder successes. Textbook level and policy level rise together. It is not a fixed textbook for a fixed exam. It selects the valuable among the solved, teaches it, and crosses the next ridge. The name climbing fits well. It is not a leap to the summit but a climb that builds one foothold at a time. This mechanism reads as the true substance of the climbing in the title.

How to read the numbers across four benchmarks and two scales

Time to look at the summit photos. With context management enabled, Iris-mini scores 82.2, 84.8, 86.9, and 52.3 on BrowseComp, BrowseComp-ZH, DeepSearchQA, and HLE in that order. Iris-pro scores 88.6, 85.1, 92.9, and 56.4 in the same order. Two scales stand side by side and four testbeds stand together. Before unpacking each figure, read the arrangement itself. Showing four testbeds rather than one, and two scales rather than one, widens the breadth of the evidence.

Recall the character of each testbed. The BrowseComp family is known for exploratory questions that must be pieced across pages. The presence of a Chinese edition invites attention, because it asks whether exploration skill carries across languages. DeepSearchQA, as the name suggests, tests deep exploration with questions that never end in one search. HLE gathers hard questions across fields and its scores run low by nature. Figures in the fifties may look modest but sit high in that context. Each testbed carries its own yardstick, so direct comparison of raw figures needs care.

Read the gap between the two scales. The small one already stands high and the large one covers above it. The small scale record alone is striking for a single agent. The large one climbs further on three testbeds while standing near the small one on the Chinese testbed. I read this as a signal that scale is not almighty. Scaling up generally lifts scores but the lift varies by testbed. Some routes gain much height and others hold level. Scale pushes skill upward, but the shape of the path is set by the testbed. Different routes gain height differently.

Do not miss the note that context management was enabled. Long explorations overflow memory easily. Read documents pile up and search histories lengthen until the policy forgets the front. Context management slows that forgetting by deciding what to keep and what to fold away. Reporting with it enabled means measuring in field configuration rather than bare form, with equipment worn. Read the figures with that equipment counted as part of the skill. Conversely it invites curiosity about the bare form record without equipment. Hopefully a later analysis will measure the contribution of the equipment separately.

The single ReAct choice and the road without side branches

All results come from a single ReAct agent, with no sub-agents and no test-time verification. The sentence is short and carries weight. The current fashion in search systems is division. A large task splits into pieces for several sub-agents and a verifier pass runs once an answer appears. Division lifts scores easily. It complicates the structure and obscures where the gain came from. Iris walks the other way. One agent connects searching, reading, and answering on its own.

The first virtue of this choice is clear credit. When scores rise, the credit flows to the policy and its training. The gain reflects learning rather than scaffolding. When scores stall, the repair site is equally clear. Before blaming role splits among sub-agents, examine the behavior of the policy. The research question sharpens. The task becomes proving learning on a simple machine rather than lifting scores with a complex one. In climbing terms, it reduces oxygen and fixed ropes and climbs on its own legs. The record gains a different worth.

The second virtue is reproduction and transfer. A single agent travels light. Many sub-agents complicate runtimes and raise call costs. Test-time verification adds measurement conditions and muddies comparison. A single structure carries none of that luggage. Bring the same policy and run it the same way. Meeting the released recipe, the effect compounds. Since the recipe describes a simple machine, following along is easier. Teams can borrow the skeleton of the idea without importing the laboratory equipment. Simplicity befriends sharing.

There are costs to name as well. A single agent loses its way more easily on long explorations. Splitting keeps each piece short, while doing everything alone lengthens the record. Perhaps that is why context management is mentioned alongside. On a solo road, packing matters. What to carry and what to drop decides success. The absence of test-time verification cuts both ways too. Measurement stays clean, yet in practice one verification pass often raises answer trust. Before moving the research choice straight into production, weigh that gap. The open question is how learning proved on a simple machine endures in messy fields.

What remains after the recipe is released

The release of the data pipeline and the training recipe will linger as long as the scores. A release is not a declaration but an invitation. Other teams will walk the same route and report which parts hold and which parts shake. How sensitive seed selection proves, how much the density of descriptive rewriting moves results, and how the two level filter thresholds behave will surface one by one. What is released is not the end of the research but the start of the conversation.

When my team examines such releases, we bring a shared set of questions. First comes cost structure. Reinforcement learning against live search improves with faster loops, and fast loops need a sturdy in-cluster judge and summarizer. How much that apparatus costs, and whether small teams can spin the same loop, remains open. Second comes breadth across languages and domains. The Chinese testbed record is welcome, yet the whole web is wider. When domains shift, link structures shift and description densities shift. Third comes reward gaming. Housing the judge inside stabilizes training, but the eyes of the judge become the sky of the policy. How the team managed the loopholes the judge might miss invites curiosity.

What can practitioners carry home right away. The most portable item is the eye for selecting questions. The yardstick that separates what memory solves from what materials solve transfers into daily team life. Building evaluation items while measuring both closed book and open material success raises data quality. The feel for blocking copying shortcuts also transfers. Before embedding proper nouns raw into a question, ask whether it still resolves when paraphrased. The feel for salvaging long runs transfers too. Before discarding a failed record whole, check whether its front holds exploration worth rescuing. Such habits harden teams at any scale.

I close with the question that lingers. How far can the search frontier climb. Larger models and better recipes will lift scores further. Yet after scores rise, what work remains. I nominate answer trust. Beyond whether the answer was right lies whether the evidence connects and whether a human can follow the connection. Followable answers outlast fast answers. What the Iris climb demonstrated is how to ascend. The next climb may demonstrate how to stay. After standing on the summit, making it trustworthy ground, fastening the ropes of evidence firmly, could continue the story. While awaiting that next report, I end this climbing log here.