Search papers, labs, and topics across Lattice.
This paper investigates the limitations of teaching models to perform reasoning tasks using a chain-of-thought approach, particularly focusing on tasks that require backtracking search. By analyzing nine reasoning tasks, the authors demonstrate that while some tasks can be effectively learned through fine-tuning, others, like cryptarithm, resist this method due to their inherent reliance on search rather than straightforward reasoning. The key finding reveals that when a task's solution is fundamentally based on search, the model cannot faithfully replicate the necessary reasoning process, leading to significant performance gaps despite high accuracy in simpler tasks.
When faced with search-dependent reasoning tasks, models struggle to learn effective chain-of-thought strategies, highlighting a critical limitation in current training paradigms.
It is tempting to assume any task solvable by a short program can be taught to a model as its chain-of-thought: write the steps out, fine-tune, and the model follows. This paper shows the assumption fails for an identifiable class of procedures. The testbed is nine reasoning tasks, each from a deterministic generator; public and hidden splits share generators, so held-out data proxies test accuracy. I reverse-engineer the generators into Python solvers, render them as chain-of-thought, and distill into a rank-<= 32 LoRA over a 30B (3.5B-active) Nemotron model. Forward-computable tasks install readily: lookup/arithmetic and an 8-bit boolean task transfer (>= 0.99 and 0.68). Cryptarithm does not: distilling its backtracking search holds at 0.01-0.07 across eleven chain-of-thought designs, RL from verifiable rewards, and self-training, even though a search solver answers 71% of instances. This is not a capability gap. The model does the arithmetic on 97-100% of lines and ranks the correct cipher in its top eight on 71%; it cannot carry the search forward as a left-to-right derivation. Fine-tuning learns the shape of a verifiable elimination step while its verdicts become unconditional templates, correct only 16-57% of the time ("verdict-as-token"). The ceiling holds across backbones from 3B to 671B and across fine-tuning and prompting; a controlled intervention isolates the cause: revealing the cipher key, which turns the derivation forward, lifts the same instances from 0.03 to 0.57. When a procedure's only solution is search over information-free structure, no faithful forward chain-of-thought exists to imitate. The task becomes learnable only by removing the search, precomputing its combinatorial core into a catalog and reducing the trace to recall plus verification; the 1st-place solution reaches Private LB 0.92 this way. What distills is memorization and verification, not search.
